2 * Copyright 1998, 2000 Marshall Kirk McKusick. All Rights Reserved.
4 * The soft updates code is derived from the appendix of a University
5 * of Michigan technical report (Gregory R. Ganger and Yale N. Patt,
6 * "Soft Updates: A Solution to the Metadata Update Problem in File
7 * Systems", CSE-TR-254-95, August 1995).
9 * Further information about soft updates can be obtained from:
11 * Marshall Kirk McKusick http://www.mckusick.com/softdep/
12 * 1614 Oxford Street mckusick@mckusick.com
13 * Berkeley, CA 94709-1608 +1-510-843-9542
16 * Redistribution and use in source and binary forms, with or without
17 * modification, are permitted provided that the following conditions
20 * 1. Redistributions of source code must retain the above copyright
21 * notice, this list of conditions and the following disclaimer.
22 * 2. Redistributions in binary form must reproduce the above copyright
23 * notice, this list of conditions and the following disclaimer in the
24 * documentation and/or other materials provided with the distribution.
26 * THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY
27 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
28 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
29 * DISCLAIMED. IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR
30 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38 * @(#)softdep.h 9.7 (McKusick) 6/21/00
42 #include <sys/queue.h>
45 * Allocation dependencies are handled with undo/redo on the in-memory
46 * copy of the data. A particular data dependency is eliminated when
47 * it is ALLCOMPLETE: that is ATTACHED, DEPCOMPLETE, and COMPLETE.
49 * The ATTACHED flag means that the data is not currently being written
52 * The UNDONE flag means that the data has been rolled back to a safe
53 * state for writing to the disk. When the I/O completes, the data is
54 * restored to its current form and the state reverts to ATTACHED.
55 * The data must be locked throughout the rollback, I/O, and roll
56 * forward so that the rolled back information is never visible to
59 * The COMPLETE flag indicates that the item has been written. For example,
60 * a dependency that requires that an inode be written will be marked
61 * COMPLETE after the inode has been written to disk.
63 * The DEPCOMPLETE flag indicates the completion of any other
64 * dependencies such as the writing of a cylinder group map has been
65 * completed. A dependency structure may be freed only when both it
66 * and its dependencies have completed and any rollbacks that are in
67 * progress have finished as indicated by the set of ALLCOMPLETE flags
70 * The two MKDIR flags indicate additional dependencies that must be done
71 * when creating a new directory. MKDIR_BODY is cleared when the directory
72 * data block containing the "." and ".." entries has been written.
73 * MKDIR_PARENT is cleared when the parent inode with the increased link
74 * count for ".." has been written. When both MKDIR flags have been
75 * cleared, the DEPCOMPLETE flag is set to indicate that the directory
76 * dependencies have been completed. The writing of the directory inode
77 * itself sets the COMPLETE flag which then allows the directory entry for
78 * the new directory to be written to disk. The RMDIR flag marks a dirrem
79 * structure as representing the removal of a directory rather than a
80 * file. When the removal dependencies are completed, additional work needs
81 * to be done* (an additional decrement of the associated inode, and a
82 * decrement of the parent inode).
84 * The DIRCHG flag marks a diradd structure as representing the changing
85 * of an existing entry rather than the addition of a new one. When
86 * the update is complete the dirrem associated with the inode for
87 * the old name must be added to the worklist to do the necessary
88 * reference count decrement.
90 * The GOINGAWAY flag indicates that the data structure is frozen from
91 * further change until its dependencies have been completed and its
92 * resources freed after which it will be discarded.
94 * The IOSTARTED flag prevents multiple calls to the I/O start routine from
95 * doing multiple rollbacks.
97 * The NEWBLOCK flag marks pagedep structures that have just been allocated,
98 * so must be claimed by the inode before all dependencies are complete.
100 * The INPROGRESS flag marks worklist structures that are still on the
101 * worklist, but are being considered for action by some process.
103 * The UFS1FMT flag indicates that the inode being processed is a ufs1 format.
105 * The EXTDATA flag indicates that the allocdirect describes an
106 * extended-attributes dependency.
108 * The ONWORKLIST flag shows whether the structure is currently linked
111 * The UNLINK* flags track the progress of updating the on-disk linked
112 * list of active but unlinked inodes. When an inode is first unlinked
113 * it is marked as UNLINKED. When its on-disk di_freelink has been
114 * written its UNLINKNEXT flags is set. When its predecessor in the
115 * list has its di_freelink pointing at us its UNLINKPREV is set.
116 * When the on-disk list can reach it from the superblock, its
117 * UNLINKONLIST flag is set. Once all of these flags are set, it
118 * is safe to let its last name be removed.
120 #define ATTACHED 0x000001
121 #define UNDONE 0x000002
122 #define COMPLETE 0x000004
123 #define DEPCOMPLETE 0x000008
124 #define MKDIR_PARENT 0x000010 /* diradd, mkdir, jaddref, jsegdep only */
125 #define MKDIR_BODY 0x000020 /* diradd, mkdir, jaddref only */
126 #define RMDIR 0x000040 /* dirrem only */
127 #define DIRCHG 0x000080 /* diradd, dirrem only */
128 #define GOINGAWAY 0x000100 /* indirdep, jremref only */
129 #define IOSTARTED 0x000200 /* inodedep, pagedep, bmsafemap only */
130 #define DELAYEDFREE 0x000400 /* allocindirect free delayed. */
131 #define NEWBLOCK 0x000800 /* pagedep, jaddref only */
132 #define INPROGRESS 0x001000 /* dirrem, freeblks, freefrag, freefile only */
133 #define UFS1FMT 0x002000 /* indirdep only */
134 #define EXTDATA 0x004000 /* allocdirect only */
135 #define ONWORKLIST 0x008000
136 #define IOWAITING 0x010000 /* Thread is waiting for IO to complete. */
137 #define ONDEPLIST 0x020000 /* Structure is on a dependency list. */
138 #define UNLINKED 0x040000 /* inodedep has been unlinked. */
139 #define UNLINKNEXT 0x080000 /* inodedep has valid di_freelink */
140 #define UNLINKPREV 0x100000 /* inodedep is pointed at in the unlink list */
141 #define UNLINKONLIST 0x200000 /* inodedep is in the unlinked list on disk */
142 #define UNLINKLINKS (UNLINKNEXT | UNLINKPREV)
144 #define ALLCOMPLETE (ATTACHED | COMPLETE | DEPCOMPLETE)
147 * The workitem queue.
149 * It is sometimes useful and/or necessary to clean up certain dependencies
150 * in the background rather than during execution of an application process
151 * or interrupt service routine. To realize this, we append dependency
152 * structures corresponding to such tasks to a "workitem" queue. In a soft
153 * updates implementation, most pending workitems should not wait for more
154 * than a couple of seconds, so the filesystem syncer process awakens once
155 * per second to process the items on the queue.
158 /* LIST_HEAD(workhead, worklist); -- declared in buf.h */
161 * Each request can be linked onto a work queue through its worklist structure.
162 * To avoid the need for a pointer to the structure itself, this structure
163 * MUST be declared FIRST in each type in which it appears! If more than one
164 * worklist is needed in the structure, then a wk_data field must be added
165 * and the macros below changed to use it.
168 LIST_ENTRY(worklist) wk_list; /* list of work requests */
169 struct mount *wk_mp; /* Mount we live in */
170 unsigned int wk_type:8, /* type of request */
171 wk_state:24; /* state flags */
173 #define WK_DATA(wk) ((void *)(wk))
174 #define WK_PAGEDEP(wk) ((struct pagedep *)(wk))
175 #define WK_INODEDEP(wk) ((struct inodedep *)(wk))
176 #define WK_BMSAFEMAP(wk) ((struct bmsafemap *)(wk))
177 #define WK_NEWBLK(wk) ((struct newblk *)(wk))
178 #define WK_ALLOCDIRECT(wk) ((struct allocdirect *)(wk))
179 #define WK_INDIRDEP(wk) ((struct indirdep *)(wk))
180 #define WK_ALLOCINDIR(wk) ((struct allocindir *)(wk))
181 #define WK_FREEFRAG(wk) ((struct freefrag *)(wk))
182 #define WK_FREEBLKS(wk) ((struct freeblks *)(wk))
183 #define WK_FREEWORK(wk) ((struct freework *)(wk))
184 #define WK_FREEFILE(wk) ((struct freefile *)(wk))
185 #define WK_DIRADD(wk) ((struct diradd *)(wk))
186 #define WK_MKDIR(wk) ((struct mkdir *)(wk))
187 #define WK_DIRREM(wk) ((struct dirrem *)(wk))
188 #define WK_NEWDIRBLK(wk) ((struct newdirblk *)(wk))
189 #define WK_JADDREF(wk) ((struct jaddref *)(wk))
190 #define WK_JREMREF(wk) ((struct jremref *)(wk))
191 #define WK_JMVREF(wk) ((struct jmvref *)(wk))
192 #define WK_JSEGDEP(wk) ((struct jsegdep *)(wk))
193 #define WK_JSEG(wk) ((struct jseg *)(wk))
194 #define WK_JNEWBLK(wk) ((struct jnewblk *)(wk))
195 #define WK_JFREEBLK(wk) ((struct jfreeblk *)(wk))
196 #define WK_FREEDEP(wk) ((struct freedep *)(wk))
197 #define WK_JFREEFRAG(wk) ((struct jfreefrag *)(wk))
198 #define WK_SBDEP(wk) ((struct sbdep *)(wk))
199 #define WK_JTRUNC(wk) ((struct jtrunc *)(wk))
200 #define WK_JFSYNC(wk) ((struct jfsync *)(wk))
203 * Various types of lists
205 LIST_HEAD(dirremhd, dirrem);
206 LIST_HEAD(diraddhd, diradd);
207 LIST_HEAD(newblkhd, newblk);
208 LIST_HEAD(inodedephd, inodedep);
209 LIST_HEAD(allocindirhd, allocindir);
210 LIST_HEAD(allocdirecthd, allocdirect);
211 TAILQ_HEAD(allocdirectlst, allocdirect);
212 LIST_HEAD(indirdephd, indirdep);
213 LIST_HEAD(jaddrefhd, jaddref);
214 LIST_HEAD(jremrefhd, jremref);
215 LIST_HEAD(jmvrefhd, jmvref);
216 LIST_HEAD(jnewblkhd, jnewblk);
217 LIST_HEAD(jblkdephd, jblkdep);
218 LIST_HEAD(freeworkhd, freework);
219 TAILQ_HEAD(freeworklst, freework);
220 TAILQ_HEAD(jseglst, jseg);
221 TAILQ_HEAD(inoreflst, inoref);
222 TAILQ_HEAD(freeblklst, freeblks);
225 * The "pagedep" structure tracks the various dependencies related to
226 * a particular directory page. If a directory page has any dependencies,
227 * it will have a pagedep linked to its associated buffer. The
228 * pd_dirremhd list holds the list of dirrem requests which decrement
229 * inode reference counts. These requests are processed after the
230 * directory page with the corresponding zero'ed entries has been
231 * written. The pd_diraddhd list maintains the list of diradd requests
232 * which cannot be committed until their corresponding inode has been
233 * written to disk. Because a directory may have many new entries
234 * being created, several lists are maintained hashed on bits of the
235 * offset of the entry into the directory page to keep the lists from
236 * getting too long. Once a new directory entry has been cleared to
237 * be written, it is moved to the pd_pendinghd list. After the new
238 * entry has been written to disk it is removed from the pd_pendinghd
239 * list, any removed operations are done, and the dependency structure
243 #define DIRADDHASH(offset) (((offset) >> 2) % DAHASHSZ)
245 struct worklist pd_list; /* page buffer */
246 # define pd_state pd_list.wk_state /* check for multiple I/O starts */
247 LIST_ENTRY(pagedep) pd_hash; /* hashed lookup */
248 ino_t pd_ino; /* associated file */
249 ufs_lbn_t pd_lbn; /* block within file */
250 struct newdirblk *pd_newdirblk; /* associated newdirblk if NEWBLOCK */
251 struct dirremhd pd_dirremhd; /* dirrem's waiting for page */
252 struct diraddhd pd_diraddhd[DAHASHSZ]; /* diradd dir entry updates */
253 struct diraddhd pd_pendinghd; /* directory entries awaiting write */
254 struct jmvrefhd pd_jmvrefhd; /* Dependent journal writes. */
258 * The "inodedep" structure tracks the set of dependencies associated
259 * with an inode. One task that it must manage is delayed operations
260 * (i.e., work requests that must be held until the inodedep's associated
261 * inode has been written to disk). Getting an inode from its incore
262 * state to the disk requires two steps to be taken by the filesystem
263 * in this order: first the inode must be copied to its disk buffer by
264 * the VOP_UPDATE operation; second the inode's buffer must be written
265 * to disk. To ensure that both operations have happened in the required
266 * order, the inodedep maintains two lists. Delayed operations are
267 * placed on the id_inowait list. When the VOP_UPDATE is done, all
268 * operations on the id_inowait list are moved to the id_bufwait list.
269 * When the buffer is written, the items on the id_bufwait list can be
270 * safely moved to the work queue to be processed. A second task of the
271 * inodedep structure is to track the status of block allocation within
272 * the inode. Each block that is allocated is represented by an
273 * "allocdirect" structure (see below). It is linked onto the id_newinoupdt
274 * list until both its contents and its allocation in the cylinder
275 * group map have been written to disk. Once these dependencies have been
276 * satisfied, it is removed from the id_newinoupdt list and any followup
277 * actions such as releasing the previous block or fragment are placed
278 * on the id_inowait list. When an inode is updated (a VOP_UPDATE is
279 * done), the "inodedep" structure is linked onto the buffer through
280 * its worklist. Thus, it will be notified when the buffer is about
281 * to be written and when it is done. At the update time, all the
282 * elements on the id_newinoupdt list are moved to the id_inoupdt list
283 * since those changes are now relevant to the copy of the inode in the
284 * buffer. Also at update time, the tasks on the id_inowait list are
285 * moved to the id_bufwait list so that they will be executed when
286 * the updated inode has been written to disk. When the buffer containing
287 * the inode is written to disk, any updates listed on the id_inoupdt
288 * list are rolled back as they are not yet safe. Following the write,
289 * the changes are once again rolled forward and any actions on the
290 * id_bufwait list are processed (since those actions are now safe).
291 * The entries on the id_inoupdt and id_newinoupdt lists must be kept
292 * sorted by logical block number to speed the calculation of the size
293 * of the rolled back inode (see explanation in initiate_write_inodeblock).
294 * When a directory entry is created, it is represented by a diradd.
295 * The diradd is added to the id_inowait list as it cannot be safely
296 * written to disk until the inode that it represents is on disk. After
297 * the inode is written, the id_bufwait list is processed and the diradd
298 * entries are moved to the id_pendinghd list where they remain until
299 * the directory block containing the name has been written to disk.
300 * The purpose of keeping the entries on the id_pendinghd list is so that
301 * the softdep_fsync function can find and push the inode's directory
302 * name(s) as part of the fsync operation for that file.
305 struct worklist id_list; /* buffer holding inode block */
306 # define id_state id_list.wk_state /* inode dependency state */
307 LIST_ENTRY(inodedep) id_hash; /* hashed lookup */
308 TAILQ_ENTRY(inodedep) id_unlinked; /* Unlinked but ref'd inodes */
309 struct fs *id_fs; /* associated filesystem */
310 ino_t id_ino; /* dependent inode */
311 nlink_t id_nlinkdelta; /* saved effective link count */
312 nlink_t id_savednlink; /* Link saved during rollback */
313 LIST_ENTRY(inodedep) id_deps; /* bmsafemap's list of inodedep's */
314 struct bmsafemap *id_bmsafemap; /* related bmsafemap (if pending) */
315 struct diradd *id_mkdiradd; /* diradd for a mkdir. */
316 struct inoreflst id_inoreflst; /* Inode reference adjustments. */
317 long id_savedextsize; /* ext size saved during rollback */
318 off_t id_savedsize; /* file size saved during rollback */
319 struct dirremhd id_dirremhd; /* Removals pending. */
320 struct workhead id_pendinghd; /* entries awaiting directory write */
321 struct workhead id_bufwait; /* operations after inode written */
322 struct workhead id_inowait; /* operations waiting inode update */
323 struct allocdirectlst id_inoupdt; /* updates before inode written */
324 struct allocdirectlst id_newinoupdt; /* updates when inode written */
325 struct allocdirectlst id_extupdt; /* extdata updates pre-inode write */
326 struct allocdirectlst id_newextupdt; /* extdata updates at ino write */
327 struct freeblklst id_freeblklst; /* List of partial truncates. */
329 struct ufs1_dinode *idu_savedino1; /* saved ufs1_dinode contents */
330 struct ufs2_dinode *idu_savedino2; /* saved ufs2_dinode contents */
333 #define id_savedino1 id_un.idu_savedino1
334 #define id_savedino2 id_un.idu_savedino2
337 * A "bmsafemap" structure maintains a list of dependency structures
338 * that depend on the update of a particular cylinder group map.
339 * It has lists for newblks, allocdirects, allocindirs, and inodedeps.
340 * It is attached to the buffer of a cylinder group block when any of
341 * these things are allocated from the cylinder group. It is freed
342 * after the cylinder group map is written and the state of its
343 * dependencies are updated with DEPCOMPLETE to indicate that it has
347 struct worklist sm_list; /* cylgrp buffer */
348 # define sm_state sm_list.wk_state
349 LIST_ENTRY(bmsafemap) sm_hash; /* Hash links. */
350 LIST_ENTRY(bmsafemap) sm_next; /* Mount list. */
352 struct buf *sm_buf; /* associated buffer */
353 struct allocdirecthd sm_allocdirecthd; /* allocdirect deps */
354 struct allocdirecthd sm_allocdirectwr; /* writing allocdirect deps */
355 struct allocindirhd sm_allocindirhd; /* allocindir deps */
356 struct allocindirhd sm_allocindirwr; /* writing allocindir deps */
357 struct inodedephd sm_inodedephd; /* inodedep deps */
358 struct inodedephd sm_inodedepwr; /* writing inodedep deps */
359 struct newblkhd sm_newblkhd; /* newblk deps */
360 struct newblkhd sm_newblkwr; /* writing newblk deps */
361 struct jaddrefhd sm_jaddrefhd; /* Pending inode allocations. */
362 struct jnewblkhd sm_jnewblkhd; /* Pending block allocations. */
363 struct workhead sm_freehd; /* Freedep deps. */
364 struct workhead sm_freewr; /* Written freedeps. */
368 * A "newblk" structure is attached to a bmsafemap structure when a block
369 * or fragment is allocated from a cylinder group. Its state is set to
370 * DEPCOMPLETE when its cylinder group map is written. It is converted to
371 * an allocdirect or allocindir allocation once the allocator calls the
372 * appropriate setup function. It will initially be linked onto a bmsafemap
373 * list. Once converted it can be linked onto the lists described for
374 * allocdirect or allocindir as described below.
377 struct worklist nb_list; /* See comment above. */
378 # define nb_state nb_list.wk_state
379 LIST_ENTRY(newblk) nb_hash; /* Hashed lookup. */
380 LIST_ENTRY(newblk) nb_deps; /* Bmsafemap's list of newblks. */
381 struct jnewblk *nb_jnewblk; /* New block journal entry. */
382 struct bmsafemap *nb_bmsafemap;/* Cylgrp dep (if pending). */
383 struct freefrag *nb_freefrag; /* Fragment to be freed (if any). */
384 struct indirdephd nb_indirdeps; /* Children indirect blocks. */
385 struct workhead nb_newdirblk; /* Dir block to notify when written. */
386 struct workhead nb_jwork; /* Journal work pending. */
387 ufs2_daddr_t nb_newblkno; /* New value of block pointer. */
391 * An "allocdirect" structure is attached to an "inodedep" when a new block
392 * or fragment is allocated and pointed to by the inode described by
393 * "inodedep". The worklist is linked to the buffer that holds the block.
394 * When the block is first allocated, it is linked to the bmsafemap
395 * structure associated with the buffer holding the cylinder group map
396 * from which it was allocated. When the cylinder group map is written
397 * to disk, ad_state has the DEPCOMPLETE flag set. When the block itself
398 * is written, the COMPLETE flag is set. Once both the cylinder group map
399 * and the data itself have been written, it is safe to write the inode
400 * that claims the block. If there was a previous fragment that had been
401 * allocated before the file was increased in size, the old fragment may
402 * be freed once the inode claiming the new block is written to disk.
403 * This ad_fragfree request is attached to the id_inowait list of the
404 * associated inodedep (pointed to by ad_inodedep) for processing after
405 * the inode is written. When a block is allocated to a directory, an
406 * fsync of a file whose name is within that block must ensure not only
407 * that the block containing the file name has been written, but also
408 * that the on-disk inode references that block. When a new directory
409 * block is created, we allocate a newdirblk structure which is linked
410 * to the associated allocdirect (on its ad_newdirblk list). When the
411 * allocdirect has been satisfied, the newdirblk structure is moved to
412 * the inodedep id_bufwait list of its directory to await the inode
413 * being written. When the inode is written, the directory entries are
414 * fully committed and can be deleted from their pagedep->id_pendinghd
415 * and inodedep->id_pendinghd lists.
418 struct newblk ad_block; /* Common block logic */
419 # define ad_state ad_block.nb_list.wk_state /* block pointer state */
420 TAILQ_ENTRY(allocdirect) ad_next; /* inodedep's list of allocdirect's */
421 struct inodedep *ad_inodedep; /* associated inodedep */
422 ufs2_daddr_t ad_oldblkno; /* old value of block pointer */
423 int ad_offset; /* Pointer offset in parent. */
424 long ad_newsize; /* size of new block */
425 long ad_oldsize; /* size of old block */
427 #define ad_newblkno ad_block.nb_newblkno
428 #define ad_freefrag ad_block.nb_freefrag
429 #define ad_newdirblk ad_block.nb_newdirblk
432 * A single "indirdep" structure manages all allocation dependencies for
433 * pointers in an indirect block. The up-to-date state of the indirect
434 * block is stored in ir_savedata. The set of pointers that may be safely
435 * written to the disk is stored in ir_safecopy. The state field is used
436 * only to track whether the buffer is currently being written (in which
437 * case it is not safe to update ir_safecopy). Ir_deplisthd contains the
438 * list of allocindir structures, one for each block that needs to be
439 * written to disk. Once the block and its bitmap allocation have been
440 * written the safecopy can be updated to reflect the allocation and the
441 * allocindir structure freed. If ir_state indicates that an I/O on the
442 * indirect block is in progress when ir_safecopy is to be updated, the
443 * update is deferred by placing the allocindir on the ir_donehd list.
444 * When the I/O on the indirect block completes, the entries on the
445 * ir_donehd list are processed by updating their corresponding ir_safecopy
446 * pointers and then freeing the allocindir structure.
449 struct worklist ir_list; /* buffer holding indirect block */
450 # define ir_state ir_list.wk_state /* indirect block pointer state */
451 LIST_ENTRY(indirdep) ir_next; /* alloc{direct,indir} list */
452 TAILQ_HEAD(, freework) ir_trunc; /* List of truncations. */
453 caddr_t ir_saveddata; /* buffer cache contents */
454 struct buf *ir_savebp; /* buffer holding safe copy */
455 struct buf *ir_bp; /* buffer holding live copy */
456 struct allocindirhd ir_completehd; /* waiting for indirdep complete */
457 struct allocindirhd ir_writehd; /* Waiting for the pointer write. */
458 struct allocindirhd ir_donehd; /* done waiting to update safecopy */
459 struct allocindirhd ir_deplisthd; /* allocindir deps for this block */
460 struct freeblks *ir_freeblks; /* Freeblks that frees this indir. */
464 * An "allocindir" structure is attached to an "indirdep" when a new block
465 * is allocated and pointed to by the indirect block described by the
466 * "indirdep". The worklist is linked to the buffer that holds the new block.
467 * When the block is first allocated, it is linked to the bmsafemap
468 * structure associated with the buffer holding the cylinder group map
469 * from which it was allocated. When the cylinder group map is written
470 * to disk, ai_state has the DEPCOMPLETE flag set. When the block itself
471 * is written, the COMPLETE flag is set. Once both the cylinder group map
472 * and the data itself have been written, it is safe to write the entry in
473 * the indirect block that claims the block; the "allocindir" dependency
474 * can then be freed as it is no longer applicable.
477 struct newblk ai_block; /* Common block area */
478 # define ai_state ai_block.nb_list.wk_state /* indirect pointer state */
479 LIST_ENTRY(allocindir) ai_next; /* indirdep's list of allocindir's */
480 struct indirdep *ai_indirdep; /* address of associated indirdep */
481 ufs2_daddr_t ai_oldblkno; /* old value of block pointer */
482 ufs_lbn_t ai_lbn; /* Logical block number. */
483 int ai_offset; /* Pointer offset in parent. */
485 #define ai_newblkno ai_block.nb_newblkno
486 #define ai_freefrag ai_block.nb_freefrag
487 #define ai_newdirblk ai_block.nb_newdirblk
490 * The allblk union is used to size the newblk structure on allocation so
491 * that it may be any one of three types.
494 struct allocindir ab_allocindir;
495 struct allocdirect ab_allocdirect;
496 struct newblk ab_newblk;
500 * A "freefrag" structure is attached to an "inodedep" when a previously
501 * allocated fragment is replaced with a larger fragment, rather than extended.
502 * The "freefrag" structure is constructed and attached when the replacement
503 * block is first allocated. It is processed after the inode claiming the
504 * bigger block that replaces it has been written to disk.
507 struct worklist ff_list; /* id_inowait or delayed worklist */
508 # define ff_state ff_list.wk_state
509 struct worklist *ff_jdep; /* Associated journal entry. */
510 struct workhead ff_jwork; /* Journal work pending. */
511 ufs2_daddr_t ff_blkno; /* fragment physical block number */
512 long ff_fragsize; /* size of fragment being deleted */
513 ino_t ff_inum; /* owning inode number */
514 enum vtype ff_vtype; /* owning inode's file type */
518 * A "freeblks" structure is attached to an "inodedep" when the
519 * corresponding file's length is reduced to zero. It records all
520 * the information needed to free the blocks of a file after its
521 * zero'ed inode has been written to disk. The actual work is done
522 * by child freework structures which are responsible for individual
523 * inode pointers while freeblks is responsible for retiring the
524 * entire operation when it is complete and holding common members.
527 struct worklist fb_list; /* id_inowait or delayed worklist */
528 # define fb_state fb_list.wk_state /* inode and dirty block state */
529 TAILQ_ENTRY(freeblks) fb_next; /* List of inode truncates. */
530 struct jblkdephd fb_jblkdephd; /* Journal entries pending */
531 struct workhead fb_freeworkhd; /* Work items pending */
532 struct workhead fb_jwork; /* Journal work pending */
533 struct vnode *fb_devvp; /* filesystem device vnode */
535 struct dquot *fb_quota[MAXQUOTAS]; /* quotas to be adjusted */
537 uint64_t fb_modrev; /* Inode revision at start of trunc. */
538 off_t fb_len; /* Length we're truncating to. */
539 ufs2_daddr_t fb_chkcnt; /* Blocks released. */
540 ino_t fb_inum; /* inode owner of blocks */
541 enum vtype fb_vtype; /* inode owner's file type */
542 uid_t fb_uid; /* uid of previous owner of blocks */
543 int fb_ref; /* Children outstanding. */
544 int fb_cgwait; /* cg writes outstanding. */
548 * A "freework" structure handles the release of a tree of blocks or a single
549 * block. Each indirect block in a tree is allocated its own freework
550 * structure so that the indirect block may be freed only when all of its
551 * children are freed. In this way we enforce the rule that an allocated
552 * block must have a valid path to a root that is journaled. Each child
553 * block acquires a reference and when the ref hits zero the parent ref
554 * is decremented. If there is no parent the freeblks ref is decremented.
557 struct worklist fw_list; /* Delayed worklist. */
558 # define fw_state fw_list.wk_state
559 LIST_ENTRY(freework) fw_segs; /* Seg list. */
560 TAILQ_ENTRY(freework) fw_next; /* Hash/Trunc list. */
561 struct jnewblk *fw_jnewblk; /* Journal entry to cancel. */
562 struct freeblks *fw_freeblks; /* Root of operation. */
563 struct freework *fw_parent; /* Parent indirect. */
564 struct indirdep *fw_indir; /* indirect block. */
565 ufs2_daddr_t fw_blkno; /* Our block #. */
566 ufs_lbn_t fw_lbn; /* Original lbn before free. */
567 uint16_t fw_frags; /* Number of frags. */
568 uint16_t fw_ref; /* Number of children out. */
569 uint16_t fw_off; /* Current working position. */
570 uint16_t fw_start; /* Start of partial truncate. */
574 * A "freedep" structure is allocated to track the completion of a bitmap
575 * write for a freework. One freedep may cover many freed blocks so long
576 * as they reside in the same cylinder group. When the cg is written
577 * the freedep decrements the ref on the freework which may permit it
578 * to be freed as well.
581 struct worklist fd_list; /* Delayed worklist. */
582 struct freework *fd_freework; /* Parent freework. */
586 * A "freefile" structure is attached to an inode when its
587 * link count is reduced to zero. It marks the inode as free in
588 * the cylinder group map after the zero'ed inode has been written
589 * to disk and any associated blocks and fragments have been freed.
592 struct worklist fx_list; /* id_inowait or delayed worklist */
593 mode_t fx_mode; /* mode of inode */
594 ino_t fx_oldinum; /* inum of the unlinked file */
595 struct vnode *fx_devvp; /* filesystem device vnode */
596 struct workhead fx_jwork; /* journal work pending. */
600 * A "diradd" structure is linked to an "inodedep" id_inowait list when a
601 * new directory entry is allocated that references the inode described
602 * by "inodedep". When the inode itself is written (either the initial
603 * allocation for new inodes or with the increased link count for
604 * existing inodes), the COMPLETE flag is set in da_state. If the entry
605 * is for a newly allocated inode, the "inodedep" structure is associated
606 * with a bmsafemap which prevents the inode from being written to disk
607 * until the cylinder group has been updated. Thus the da_state COMPLETE
608 * flag cannot be set until the inode bitmap dependency has been removed.
609 * When creating a new file, it is safe to write the directory entry that
610 * claims the inode once the referenced inode has been written. Since
611 * writing the inode clears the bitmap dependencies, the DEPCOMPLETE flag
612 * in the diradd can be set unconditionally when creating a file. When
613 * creating a directory, there are two additional dependencies described by
614 * mkdir structures (see their description below). When these dependencies
615 * are resolved the DEPCOMPLETE flag is set in the diradd structure.
616 * If there are multiple links created to the same inode, there will be
617 * a separate diradd structure created for each link. The diradd is
618 * linked onto the pg_diraddhd list of the pagedep for the directory
619 * page that contains the entry. When a directory page is written,
620 * the pg_diraddhd list is traversed to rollback any entries that are
621 * not yet ready to be written to disk. If a directory entry is being
622 * changed (by rename) rather than added, the DIRCHG flag is set and
623 * the da_previous entry points to the entry that will be "removed"
624 * once the new entry has been committed. During rollback, entries
625 * with da_previous are replaced with the previous inode number rather
628 * The overlaying of da_pagedep and da_previous is done to keep the
629 * structure down. If a da_previous entry is present, the pointer to its
630 * pagedep is available in the associated dirrem entry. If the DIRCHG flag
631 * is set, the da_previous entry is valid; if not set the da_pagedep entry
632 * is valid. The DIRCHG flag never changes; it is set when the structure
633 * is created if appropriate and is never cleared.
636 struct worklist da_list; /* id_inowait or id_pendinghd list */
637 # define da_state da_list.wk_state /* state of the new directory entry */
638 LIST_ENTRY(diradd) da_pdlist; /* pagedep holding directory block */
639 doff_t da_offset; /* offset of new dir entry in dir blk */
640 ino_t da_newinum; /* inode number for the new dir entry */
642 struct dirrem *dau_previous; /* entry being replaced in dir change */
643 struct pagedep *dau_pagedep; /* pagedep dependency for addition */
645 struct workhead da_jwork; /* Journal work awaiting completion. */
647 #define da_previous da_un.dau_previous
648 #define da_pagedep da_un.dau_pagedep
651 * Two "mkdir" structures are needed to track the additional dependencies
652 * associated with creating a new directory entry. Normally a directory
653 * addition can be committed as soon as the newly referenced inode has been
654 * written to disk with its increased link count. When a directory is
655 * created there are two additional dependencies: writing the directory
656 * data block containing the "." and ".." entries (MKDIR_BODY) and writing
657 * the parent inode with the increased link count for ".." (MKDIR_PARENT).
658 * These additional dependencies are tracked by two mkdir structures that
659 * reference the associated "diradd" structure. When they have completed,
660 * they set the DEPCOMPLETE flag on the diradd so that it knows that its
661 * extra dependencies have been completed. The md_state field is used only
662 * to identify which type of dependency the mkdir structure is tracking.
663 * It is not used in the mainline code for any purpose other than consistency
664 * checking. All the mkdir structures in the system are linked together on
665 * a list. This list is needed so that a diradd can find its associated
666 * mkdir structures and deallocate them if it is prematurely freed (as for
667 * example if a mkdir is immediately followed by a rmdir of the same directory).
668 * Here, the free of the diradd must traverse the list to find the associated
669 * mkdir structures that reference it. The deletion would be faster if the
670 * diradd structure were simply augmented to have two pointers that referenced
671 * the associated mkdir's. However, this would increase the size of the diradd
672 * structure to speed a very infrequent operation.
675 struct worklist md_list; /* id_inowait or buffer holding dir */
676 # define md_state md_list.wk_state /* type: MKDIR_PARENT or MKDIR_BODY */
677 struct diradd *md_diradd; /* associated diradd */
678 struct jaddref *md_jaddref; /* dependent jaddref. */
679 struct buf *md_buf; /* MKDIR_BODY: buffer holding dir */
680 LIST_ENTRY(mkdir) md_mkdirs; /* list of all mkdirs */
684 * A "dirrem" structure describes an operation to decrement the link
685 * count on an inode. The dirrem structure is attached to the pg_dirremhd
686 * list of the pagedep for the directory page that contains the entry.
687 * It is processed after the directory page with the deleted entry has
688 * been written to disk.
691 struct worklist dm_list; /* delayed worklist */
692 # define dm_state dm_list.wk_state /* state of the old directory entry */
693 LIST_ENTRY(dirrem) dm_next; /* pagedep's list of dirrem's */
694 LIST_ENTRY(dirrem) dm_inonext; /* inodedep's list of dirrem's */
695 struct jremrefhd dm_jremrefhd; /* Pending remove reference deps. */
696 ino_t dm_oldinum; /* inum of the removed dir entry */
697 doff_t dm_offset; /* offset of removed dir entry in blk */
699 struct pagedep *dmu_pagedep; /* pagedep dependency for remove */
700 ino_t dmu_dirinum; /* parent inode number (for rmdir) */
702 struct workhead dm_jwork; /* Journal work awaiting completion. */
704 #define dm_pagedep dm_un.dmu_pagedep
705 #define dm_dirinum dm_un.dmu_dirinum
708 * A "newdirblk" structure tracks the progress of a newly allocated
709 * directory block from its creation until it is claimed by its on-disk
710 * inode. When a block is allocated to a directory, an fsync of a file
711 * whose name is within that block must ensure not only that the block
712 * containing the file name has been written, but also that the on-disk
713 * inode references that block. When a new directory block is created,
714 * we allocate a newdirblk structure which is linked to the associated
715 * allocdirect (on its ad_newdirblk list). When the allocdirect has been
716 * satisfied, the newdirblk structure is moved to the inodedep id_bufwait
717 * list of its directory to await the inode being written. When the inode
718 * is written, the directory entries are fully committed and can be
719 * deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd
720 * lists. Note that we could track directory blocks allocated to indirect
721 * blocks using a similar scheme with the allocindir structures. Rather
722 * than adding this level of complexity, we simply write those newly
723 * allocated indirect blocks synchronously as such allocations are rare.
724 * In the case of a new directory the . and .. links are tracked with
725 * a mkdir rather than a pagedep. In this case we track the mkdir
726 * so it can be released when it is written. A workhead is used
727 * to simplify canceling a mkdir that is removed by a subsequent dirrem.
730 struct worklist db_list; /* id_inowait or pg_newdirblk */
731 # define db_state db_list.wk_state
732 struct pagedep *db_pagedep; /* associated pagedep */
733 struct workhead db_mkdir;
737 * The inoref structure holds the elements common to jaddref and jremref
738 * so they may easily be queued in-order on the inodedep.
741 struct worklist if_list; /* Journal pending or jseg entries. */
742 # define if_state if_list.wk_state
743 TAILQ_ENTRY(inoref) if_deps; /* Links for inodedep. */
744 struct jsegdep *if_jsegdep; /* Will track our journal record. */
745 off_t if_diroff; /* Directory offset. */
746 ino_t if_ino; /* Inode number. */
747 ino_t if_parent; /* Parent inode number. */
748 nlink_t if_nlink; /* nlink before addition. */
749 uint16_t if_mode; /* File mode, needed for IFMT. */
753 * A "jaddref" structure tracks a new reference (link count) on an inode
754 * and prevents the link count increase and bitmap allocation until a
755 * journal entry can be written. Once the journal entry is written,
756 * the inode is put on the pendinghd of the bmsafemap and a diradd or
757 * mkdir entry is placed on the bufwait list of the inode. The DEPCOMPLETE
758 * flag is used to indicate that all of the required information for writing
759 * the journal entry is present. MKDIR_BODY and MKDIR_PARENT are used to
760 * differentiate . and .. links from regular file names. NEWBLOCK indicates
761 * a bitmap is still pending. If a new reference is canceled by a delete
762 * prior to writing the journal the jaddref write is canceled and the
763 * structure persists to prevent any disk-visible changes until it is
764 * ultimately released when the file is freed or the link is dropped again.
767 struct inoref ja_ref; /* see inoref above. */
768 # define ja_list ja_ref.if_list /* Jrnl pending, id_inowait, dm_jwork.*/
769 # define ja_state ja_ref.if_list.wk_state
770 LIST_ENTRY(jaddref) ja_bmdeps; /* Links for bmsafemap. */
772 struct diradd *jau_diradd; /* Pending diradd. */
773 struct mkdir *jau_mkdir; /* MKDIR_{PARENT,BODY} */
776 #define ja_diradd ja_un.jau_diradd
777 #define ja_mkdir ja_un.jau_mkdir
778 #define ja_diroff ja_ref.if_diroff
779 #define ja_ino ja_ref.if_ino
780 #define ja_parent ja_ref.if_parent
781 #define ja_mode ja_ref.if_mode
784 * A "jremref" structure tracks a removed reference (unlink) on an
785 * inode and prevents the directory remove from proceeding until the
786 * journal entry is written. Once the journal has been written the remove
787 * may proceed as normal.
790 struct inoref jr_ref; /* see inoref above. */
791 # define jr_list jr_ref.if_list /* Linked to softdep_journal_pending. */
792 # define jr_state jr_ref.if_list.wk_state
793 LIST_ENTRY(jremref) jr_deps; /* Links for dirrem. */
794 struct dirrem *jr_dirrem; /* Back pointer to dirrem. */
798 * A "jmvref" structure tracks a name relocations within the same
799 * directory block that occur as a result of directory compaction.
800 * It prevents the updated directory entry from being written to disk
801 * until the journal entry is written. Once the journal has been
802 * written the compacted directory may be written to disk.
805 struct worklist jm_list; /* Linked to softdep_journal_pending. */
806 LIST_ENTRY(jmvref) jm_deps; /* Jmvref on pagedep. */
807 struct pagedep *jm_pagedep; /* Back pointer to pagedep. */
808 ino_t jm_parent; /* Containing directory inode number. */
809 ino_t jm_ino; /* Inode number of our entry. */
810 off_t jm_oldoff; /* Our old offset in directory. */
811 off_t jm_newoff; /* Our new offset in directory. */
815 * A "jnewblk" structure tracks a newly allocated block or fragment and
816 * prevents the direct or indirect block pointer as well as the cg bitmap
817 * from being written until it is logged. After it is logged the jsegdep
818 * is attached to the allocdirect or allocindir until the operation is
819 * completed or reverted. If the operation is reverted prior to the journal
820 * write the jnewblk structure is maintained to prevent the bitmaps from
821 * reaching the disk. Ultimately the jnewblk structure will be passed
822 * to the free routine as the in memory cg is modified back to the free
823 * state at which time it can be released. It may be held on any of the
824 * fx_jwork, fw_jwork, fb_jwork, ff_jwork, nb_jwork, or ir_jwork lists.
827 struct worklist jn_list; /* See lists above. */
828 # define jn_state jn_list.wk_state
829 struct jsegdep *jn_jsegdep; /* Will track our journal record. */
830 LIST_ENTRY(jnewblk) jn_deps; /* Jnewblks on sm_jnewblkhd. */
831 struct worklist *jn_dep; /* Dependency to ref completed seg. */
832 ufs_lbn_t jn_lbn; /* Lbn to which allocated. */
833 ufs2_daddr_t jn_blkno; /* Blkno allocated */
834 ino_t jn_ino; /* Ino to which allocated. */
835 int jn_oldfrags; /* Previous fragments when extended. */
836 int jn_frags; /* Number of fragments. */
840 * A "jblkdep" structure tracks jfreeblk and jtrunc records attached to a
841 * freeblks structure.
844 struct worklist jb_list; /* For softdep journal pending. */
845 struct jsegdep *jb_jsegdep; /* Reference to the jseg. */
846 struct freeblks *jb_freeblks; /* Back pointer to freeblks. */
847 LIST_ENTRY(jblkdep) jb_deps; /* Dep list on freeblks. */
852 * A "jfreeblk" structure tracks the journal write for freeing a block
853 * or tree of blocks. The block pointer must not be cleared in the inode
854 * or indirect prior to the jfreeblk being written to the journal.
857 struct jblkdep jf_dep; /* freeblks linkage. */
858 ufs_lbn_t jf_lbn; /* Lbn from which blocks freed. */
859 ufs2_daddr_t jf_blkno; /* Blkno being freed. */
860 ino_t jf_ino; /* Ino from which blocks freed. */
861 int jf_frags; /* Number of frags being freed. */
865 * A "jfreefrag" tracks the freeing of a single block when a fragment is
866 * extended or an indirect page is replaced. It is not part of a larger
867 * freeblks operation.
870 struct worklist fr_list; /* Linked to softdep_journal_pending. */
871 # define fr_state fr_list.wk_state
872 struct jsegdep *fr_jsegdep; /* Will track our journal record. */
873 struct freefrag *fr_freefrag; /* Back pointer to freefrag. */
874 ufs_lbn_t fr_lbn; /* Lbn from which frag freed. */
875 ufs2_daddr_t fr_blkno; /* Blkno being freed. */
876 ino_t fr_ino; /* Ino from which frag freed. */
877 int fr_frags; /* Size of frag being freed. */
881 * A "jtrunc" journals the intent to truncate an inode's data or extent area.
884 struct jblkdep jt_dep; /* freeblks linkage. */
885 off_t jt_size; /* Final file size. */
886 int jt_extsize; /* Final extent size. */
887 ino_t jt_ino; /* Ino being truncated. */
891 * A "jfsync" journals the completion of an fsync which invalidates earlier
892 * jtrunc records in the journal.
895 struct worklist jfs_list; /* For softdep journal pending. */
896 off_t jfs_size; /* Sync file size. */
897 int jfs_extsize; /* Sync extent size. */
898 ino_t jfs_ino; /* ino being synced. */
902 * A "jsegdep" structure tracks a single reference to a written journal
903 * segment so the journal space can be reclaimed when all dependencies
904 * have been written. It can hang off of id_inowait, dm_jwork, da_jwork,
905 * nb_jwork, ff_jwork, or fb_jwork lists.
908 struct worklist jd_list; /* See above for lists. */
909 # define jd_state jd_list.wk_state
910 struct jseg *jd_seg; /* Our journal record. */
914 * A "jseg" structure contains all of the journal records written in a
915 * single disk write. The jaddref and jremref structures are linked into
916 * js_entries so thay may be completed when the write completes. The
917 * js_entries also include the write dependency structures: jmvref,
918 * jnewblk, jfreeblk, jfreefrag, and jtrunc. The js_refs field counts
919 * the number of entries on the js_entries list. Thus there is a single
920 * jseg entry to describe each journal write.
923 struct worklist js_list; /* b_deps link for journal */
924 # define js_state js_list.wk_state
925 struct workhead js_entries; /* Entries awaiting write */
926 LIST_HEAD(, freework) js_indirs;/* List of indirects in this seg. */
927 TAILQ_ENTRY(jseg) js_next; /* List of all unfinished segments. */
928 struct jblocks *js_jblocks; /* Back pointer to block/seg list */
929 struct buf *js_buf; /* Buffer while unwritten */
930 uint64_t js_seq; /* Journal record sequence number. */
931 uint64_t js_oldseq; /* Oldest valid sequence number. */
932 int js_size; /* Size of journal record in bytes. */
933 int js_cnt; /* Total items allocated. */
934 int js_refs; /* Count of js_entries items. */
938 * A 'sbdep' structure tracks the head of the free inode list and
939 * superblock writes. This makes sure the superblock is always pointing at
940 * the first possible unlinked inode for the suj recovery process. If a
941 * block write completes and we discover a new head is available the buf
942 * is dirtied and the dep is kept. See the description of the UNLINK*
943 * flags above for more details.
946 struct worklist sb_list; /* b_dep linkage */
947 struct fs *sb_fs; /* Filesystem pointer within buf. */
948 struct ufsmount *sb_ump; /* Our mount structure */